oct-diagn-semi-supervised

Semi-supervised classification of Retinal OCT images

Project is based on:

• Cortex - wrapper around Pytorch
• MlFlow - experiments tracking

ArXiv: arxiv.org/abs/2010.12316

Presentation: KDAH-CIKM-2020-presentation.pdf

Installation

1. Make sure, you have Python 3.7
2. Create a virtual environment:

   pip install virtualenv
   virtualenv venv
   source venv/bin/activate

3. pip3 install -r requirements.txt
4. Install dev version of cortex:

   git clone https://github.com/rdevon/cortex
   cd cortex
   git checkout dev
   pip3 install -e .

MlFlow Server

To run your own MlFlow Server with experiments: mlflow server --default-artifact-root='mlruns/'

Running scrips

Data download

To download UCSD dataset from Kaggle set up your Kaggle credentials and run: PYTHONPATH=. python3 src/data/download.py

Simple training

Fully-supervised model: PYTHONPATH=. python3 src/models/full_supervised/main.py

--d.sources ImageFolder 
--name OCT2017 
--data_args.image_size "(512, 512)" 
--data_args.random_resize_crop "{'size': (256, 256), 'scale': (0.5, 1.0)}" 
--data_args.image_size_test "(256, 256)" 
--t.epochs 25 
--d.shuffle True 
--data_args.normalize True 
--data_args.flip 
--o.learning_rate 0.001 
--data_args.split_labelled_and_unlabelled 
--v.viz_off 
--log_to_mlflow True

For Transfer Learning add three arguments:

--data_args.n_labels 200 
--pretrained 
--freeze_layers True

MixMatch: PYTHONPATH=. python3 src/models/mix_match/main.py

--d.sources ImageFolder 
--name OCT2017 
--data_args.image_size "(512, 512)" 
--data_args.random_resize_crop "{'size': (256, 256), 'scale': (0.5, 1.0)}" 
--data_args.image_size_test "(256, 256)" 
--t.epochs 1000 
--d.shuffle True 
--data_args.n_labels 200 
--data_args.normalize True 
--data_args.split_labelled_and_unlabelled 
--data_args.flip 
--lambda_u 25 
--log_to_mlflow True 
--v.viz_off

FixMatch: PYTHONPATH=. python3 src/models/fix_match/main.py

--d.sources ImageFolder 
--name OCT2017 
--data_args.image_size "(512, 512)" 
--data_args.random_resize_crop "{'size': (256, 256), 'scale': (0.5, 1.0)}" 
--data_args.image_size_test "(256, 256)" 
--t.epochs 2000 
--d.shuffle True 
--data_args.n_labels 200 
--data_args.normalize True 
--data_args.split_labelled_and_unlabelled 
--data_args.flip 
--log_to_mlflow True 
--data_args.mu 4 
--data_args.fix_match_transform 
--v.viz_off

Hyperparameter search / Varying number of labels

Set-up search grid dict in src/<model_name>/varying_number_of_labels.py and run PYTHONPATH=. python3 src/models/<model_name>/varying_number_of_labels.py

--d.sources ImageFolder 
--name OCT2017 
--type_of_run varying_number_of_labels / hparam_search
--data_args.image_size "(512, 512)" 
--data_args.random_resize_crop "{'size': (256, 256), 'scale': (0.5, 1.0)}" 
--data_args.image_size_test "(256, 256)" 
--d.shuffle True 
--data_args.normalize True 
--data_args.split_labelled_and_unlabelled 
--data_args.flip 
--log_to_mlflow True 

# Optional for Transfer Learning
--pretrained

# Optional for Fix-Match
--data_args.fix_match_transform

Project Organization

├── LICENSE
├── README.md 
├── data
│   └── OCT2017                     <- Dataset, loaded from https://www.kaggle.com/paultimothymooney/kermany2018, see src/data/download.py
|       ├── test      
|       ├── train     
│       └── val       
│
├── notebooks                       <- Jupyter notebooks
|   ├── 00_dataset_statistics.ipynb <- OCT2017 exploratory data analysis
|   └── 01_experiments.ipynb        <- Plots/graphs for experimental results  
│
├── reports                         <- Generated analysis as HTML, PDF, LaTeX, etc.
│   └── figures                     <- Generated graphics and figures to be used in reporting
│
├── requirements.txt                <- The requirements file for reproducing the analysis environment
│
├── setup.py                        <- makes project pip installable (pip install -e .) so src can be imported
│
└── src                             <- Source code for use in this project
    ├── data                        <- Scripts to download and manipulate data
    │   ├── dataset_plugins.py      <- Modification of default cortex DataLoaders for SSL
    │   ├── download.py             <- Script for data download
    │   ├── rand_augment.py         <- RandAugment functions
    │   └── transforms.py           <- Addiotional augmentations
    │
    └── models                      <- Scripts to train models and then use trained models to make predictions
        ├── fix_match               <- FixMatch
        |   ├── controller.py               <- Cortex controller 
        |   ├── main.py                     <- Train script 
        |   ├── utils.py            
        |   └── varying_number_of_labels.py <- Hyperparameter search / varying number of labels runs
        |   
        ├── full_supervised         <- Fully Supervised / Transfer Learning models
        |   ├── ...                          <- Same structure, as in src/models/fix_match
        |
        ├── mix_match               <- MixMatch
        |   ├── ...                         <- Same structure, as in src/models/fix_match
        |
        ├── utils.py
        └── wideresnet.py           <- Wide-Res-Net backbone

---

Project based on the cookiecutter data science project template. #cookiecutterdatascience